AI Alignment Amplifies the Role of Race, Gender, and Disability in Hiring Decisions

URL SCAN: arxiv.org/abs/2605.13866
FIRST LINE: "Humans increasingly delegate decisions to language models, yet whether these systems reproduce or reshape human patterns of discrimination remains unclear."

THE DISSECTION

This paper is a forensic snapshot of a machine labor market already in partial operation. It is not a policy discussion. It is not a warning. It is a system audit of how AI has begun adjudicating human economic participation—and the results are not ambiguous.

The study covers 27 models and 177 occupations. The findings:

• Female candidates receive preferential treatment relative to comparable male candidates.
• Black candidates receive preferential treatment relative to comparable white candidates.
• Disabled candidates receive systematic disadvantage.

Post-training alignment—RLHF, DPO, and similar "safety" interventions—is the primary engine. Pre-trained models show modest disparities; aligned models show amplified directional effects, sometimes 3x pre-trained levels. Alignment does not flatten bias. It redistributes it along different vectors.

THE CORE FALLACY

The framing problem here is massive. The paper is titled as though the question is whether AI reproduces or reshapes discrimination. The honest title would be: "AI Has Already Operationalized a New Discrimination Regime That Is Not Responsive to Human Preference."

The hidden assumption throughout is that the human hiring market—which the paper treats as the baseline—is the thing being compared. But humans built these models, humans aligned these models, and the models are now making decisions humans have delegated and can no longer meaningfully review at scale. The baseline is not a stable reference point. It is a system in the process of being displaced.

The paper also assumes the discrimination is a bug to be fixed. Under the Discontinuity Thesis, the relevant question is not "is this fair?" but "who benefits from this discrimination operating at scale, and does it accelerate or decelerate the structural displacement of human labor?"

THE VERDICT

This paper is a data dump from the early automation wave. What it documents is not a hypothetical future. It is a present condition. AI systems are already:

1. Operationalizing group-level preferences at scale, without individual accountability.
2. Rewriting the employment market's discrimination topology—flipping racial effects, amplifying gender effects, creating a new disabled penalty.
3. Delivering these outcomes through alignment processes controlled by a small number of labs, not through market competition or democratic deliberation.

The 325-330% amplification of female/Black advantages and 171% amplification of disabled disadvantage is not noise. It is signal. It tells you that the alignment layer—the process most celebrated as making AI "safe"—is the mechanism through which economic selection is being redistributed. The labs are not neutral. The process is not transparent. The outcome is not reversible through individual hiring decisions.

The paper's observation that "absence of qualification signals harms marginalised groups more, particularly for disabled candidates" is the most structurally important finding in the document. It means that as qualification signals become more ambiguous—because AI-generated content blurs credentials, credentials themselves become devalued, and human-documented experience becomes scarcer—disabled candidates suffer more. The paper identifies this asymmetry but does not draw the structural conclusion: this is a direct mechanism of productive participation collapse for a specific population under AI transition conditions.

THE SOCIAL FUNCTION

This paper is partial truth with institutional restraint. It provides the empirical foundation but refuses to name the systemic implication. It is the research equivalent of documenting that the new factory chimney is poisoning the water supply, while presenting the finding as a study of local groundwater chemistry. The data is important. The framing is conservative. The social function is to produce a document that can be cited by people who do not want to act on what it says.

IMPLICATION UNDER DISCONTINUITY THESIS

The DT lens identifies this as a lag-phase fragmentation of the employment system that accelerates P3 (Productive Participation Collapse):

• AI hiring systems are already selecting based on demographic vectors that differ from historical human hiring patterns.
• Disabled candidates face compounded exclusion: AI-generated credential inflation and experience-devaluation hit them hardest.
• Alignment-driven selection effects mean the bias is not in the base model but in the deliberate human choices about what "good" looks like—making it structural and intentional, not emergent.
• As these systems scale, human preference data becomes increasingly irrelevant to actual hiring outcomes.

The paper is a field report from the economic transition. The transition is already underway.